Fix SPECIAL behaviour in all-NaN reductions

Fix REDUCE_MIN, REDUCE_MAX, and MAX_POOL2D issue where reductions in
which all input values where NaNs would wrongly yield the initialisation
value when the nan_mode was set to IGNORE.

Add utilities getApplyMinPadding and getApplyMaxPadding to standardise
the way we get padding/initialisation values and updates MAX_POOL2D and
ARGMAX to use those.

Add unit tests to catch regressions and other wrong behaviour for
FP_SPECIAL REDUCE_MIN and REDUCE_MAX.

Signed-off-by: Ian Tayler Lessa <ian.taylerlessa@arm.com>
Change-Id: I9644cc82c5b7390186f80dc3dfe14c5e062eaad7